AIST has published a video showing the HRP-3 Promet Mk.II, one of their full-size humanoid robots, grasping a bottle using its multi-fingered hand based on visual recognition. Despite having three cameras in its head, only one is required to recognize the object. The grasp planning uses simple rectangular parallelepiped models (which are basically cubes) to determine the object’s size and position in space. The demonstration doesn’t show how the robot would handle objects of varying size or material (e.g. fabric), so it’s unclear how robust this particular method is compared to others.
[source: AIST Channel @ YouTube]